mirror of
https://github.com/NousResearch/atropos.git
synced 2026-05-01 17:45:16 +00:00
feef039cfd
Add EnsembleReward to atroposlib/envs/reward_fns/ with: - Multiple aggregation strategies: mean, median, min, majority_vote - Krippendorff's alpha inter-rater reliability metric - Per-item disagreement tracking for reward hacking detection - Full integration with RewardRegistry 17/17 tests passing.
31 lines
1.1 KiB
Python
31 lines
1.1 KiB
Python
"""
|
|
Reward functions for evaluating model outputs in various environments.
|
|
|
|
This module provides a framework for creating, composing, and applying reward functions
|
|
to evaluate model outputs. Reward functions can be used for both dataset environments
|
|
and online/gymnasium environments.
|
|
|
|
Key components:
|
|
- RewardFunction: Abstract base class for all reward functions
|
|
- RewardRegistry: Registry for registering and loading reward functions
|
|
- CombinedReward: Meta reward function that combines multiple reward functions
|
|
|
|
Usage:
|
|
# Define a reward function
|
|
@registry.register
|
|
class MyReward(RewardFunction):
|
|
def compute(self, completions, **kwargs):
|
|
# Implementation
|
|
return [score for completion in completions]
|
|
|
|
# Create and use a reward function
|
|
reward_fn = registry.create("my_reward", weight=1.5)
|
|
scores = reward_fn(completions, **kwargs)
|
|
"""
|
|
|
|
from .combined_reward import CombinedReward
|
|
from .ensemble_reward import EnsembleReward
|
|
from .registry import registry
|
|
from .reward_function import RewardFunction
|
|
|
|
__all__ = ["RewardFunction", "registry", "CombinedReward", "EnsembleReward"]
|